On 1/17/24 16:42, Jens Axboe wrote:
On 1/17/24 5:38 PM, Bart Van Assche wrote:
On 1/17/24 10:43, Jens Axboe wrote:
Do we care? Maybe not, if we accept that an IO scheduler is just for
"slower devices". But let's not go around spouting some 200K number as
if it's gospel, when it depends on so many factors like IO workload,
system used, etc.
I've never seen more than 200K IOPS in a single-threaded test. Since
your tests report higher IOPS numbers, I assume that you are submitting
I/O from multiple CPU cores at the same time.
Single core, using mq-deadline (with the poc patch, but number without
you can already find in a previous reply):
axboe@7950x ~/g/fio (master)> cat /sys/block/nvme0n1/queue/scheduler
none [mq-deadline]
axboe@7950x ~/g/fio (master)> sudo t/io_uring -p1 -d128 -b512 -s32 -c32 -F1 -B1 -R1 -X1 -n1 /dev/nvme0n1
submitter=0, tid=1957, file=/dev/nvme0n1, node=-1
polled=1, fixedbufs=1/0, register_files=1, buffered=0, QD=128
Engine=io_uring, sq_ring=128, cq_ring=128
IOPS=5.10M, BW=2.49GiB/s, IOS/call=32/31
IOPS=5.10M, BW=2.49GiB/s, IOS/call=32/32
IOPS=5.10M, BW=2.49GiB/s, IOS/call=31/31
Using non-polled IO, the number is around 4M.
A correction: my tests ran with 72 fio jobs instead of 1. I used
fio + io_uring + null_blk in my tests. I see about 1100 K IOPS with
a single fio job and about 150 K IOPS with 72 fio jobs. This shows
how I measured mq-deadline performance:
modprobe null_blk
fio --bs=4096 --group_reporting=1 --gtod_reduce=1 --invalidate=1 \
--ioengine=io_uring --ioscheduler=mq-deadline --norandommap \
--runtime=60 --rw=randread --thread --time_based=1 --buffered=0 \
--numjobs=72 --iodepth=128 --iodepth_batch_submit=64 \
--iodepth_batch_complete=64 --name=/dev/nullb0 --filename=/dev/nullb0
Thanks,
Bart.
